The reason your best AI use cases are still in a pilot is rarely technical. It is that the data they need is the data legal will not let leave. So teams either ship something weaker on redacted inputs, or wait.
What it changes for the business
- The sensitive use cases become approvable. The model gets a request shaped exactly like the real one, and the real values never leave your boundary.
- Your safety filtering stops being a second disclosure. Content inspection happens without shipping your text to somebody else's classifier to be read.
- Answers still make sense. Substituted values keep the shape of the original, so the model reasons normally and your application gets the real values back.
- The data-processing conversation gets shorter. There is less to argue about when the provider never receives the material in question.
How it works, briefly
Sensitive values are detected and substituted before the request is dispatched, and restored on the way back, including on streamed responses. Anything that inspects the content sees the substitutes. The mapping is never stored.
| How this is handled elsewhere | How Squidder does it |
|---|---|
| Send prompts to a hosted filter to be inspected | Nothing sensitive is sent to be inspected |
| Redact before sending, and lose the answer quality | Substitute, so the answer still works |
| Trust a retention promise and a data-processing agreement | The data was never transmitted |
| One vendor's safety service, one vendor's models | Works the same across every provider you use |
Where it fits, and where it does not
Substitution acts on values that can be identified, which is why the response ladder does not end there. For a destination where nothing unrecognised should leave at all, the call can be refused outright, and that is set per destination so you can be strict where it matters.
Where to start
Take the use case your privacy review stopped. Turn substitution on for it, run your own evaluation set, and compare the answers to the unprotected run.